Professional Documents
Culture Documents
Recommendation To Groups of Users Using The Singularities Concept
Recommendation To Groups of Users Using The Singularities Concept
Recommendation To Groups of Users Using The Singularities Concept
7, 2018.
Digital Object Identifier 10.1109/ACCESS.2018.2853107
ABSTRACT Recommendation to a group of users is a big challenge for collaborative filtering. The
recommendations to groups of users arise from the convenience of being able to recommend a group of users
about products or services that satisfy the entire group. In this paper, we propose the similarity measure
SMGU, tailored for collaborative filtering recommendations to groups of users. This similarity measure
combines both numerical and non-numerical information. Numerical information is weighted attending
to the rating singularity of the group members. This paper focuses on the assumption that the singularity
of the ratings cast by the users of the group is relevant information for finding suitable neighbors. For
each item, we consider that a rating is singular for a group or for a user when that rating is different
from the majority of the rating cast by the other users. Non-numerical structural information can be
considered as valuable to match group preferences with neighbors preferences. Experiments have been run
using open recommender systems data sets. Compared with representative baselines, results show accuracy
improvements when the proposed method is used. Additionally, this paper provides a section devoted to
the experiments reproducibility issue. Finally, this paper opens opportunities to face new challenges in the
recommendation to a group of users: explanation of recommendations, determination of reliability measures,
and improvement of accuracy, novelty, and diversity results.
INDEX TERMS Recommendation to groups, group of users, collaborative filtering, recommender systems,
singularity.
2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 39745
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
F. Ortega et al.: Recommendation to Groups of Users Using the Singularities Concept
recommend to active users based on the preferences of their from the MF approach, although their use is not extended:
closest social network. The main problem with this approach fuzzy approaches [14], evolutionary algorithms (ants, swarm,
is that most of the existing datasets do not contain enough etc.) [15], Bayesian methods [16], clustering [17], etc.
social information. Finally, context-aware RS [7], [8] are
usually associated with the Internet of Things (IoT), where B. RECOMMENDATION TO GROUPS OF USERS
context information is collected: GPS coordinates, RFID While traditional recommendations are individual (they are
information, credit card data, etc. made for each active user), this paper focuses on recommen-
Collaborative Filtering (CF) RS [1], [9] usually offer the dations to groups of users. The recommendations to groups of
best recommendation results. Their operation is as follows: users arise from the convenience of being able to recommend
the active user is recommended items that have not been con- a group of users about products or services that satisfy the
sumed and that have been positively rated by users who have entire group. A classic example is the recommendation of
preferences similar to those of the active user. That is, infor- a movie to a family or a group of friends. Sometimes the
mation is extracted from all existing users (hundreds of thou- design of the recommendation to groups method requires
sands or millions) and based on that information, the active the use of an aggregation policy [18]. The most commonly
user is recommended. Normally, information is structured used policies [19], [20] are: a) The last misery policy, which
as a matrix that stores the preferences (explicit or implicit) avoids recommendations that satisfy a majority, but are not
of each of the users about the set of items. These matrices satisfactory for a minority, and b) The weighted aggregation
(efficiently saved in datasets) are enormously sparse because policy, where it is attempted to maximize overall satisfaction,
a typical user has only been able to consume or rating a very without taking into account variations in the satisfaction of
small subset of the set of available items (thousands or tens the group’s components.
of thousands). Although the recommendation to groups of users may
There is a wide variety of approaches to extract the most seem like a simple generalization of the recommendation to
relevant information from the sparse collaborative filtering individual users, it is more complex: a) When a memory-
matrices. The traditional approach was the KNN algorithm based approach is chosen, there are several possibilities to
(K Nearest Neighbors) [10], [11], where the most similar approach the problem: You have to choose between making
K users (neighborhood) are searched for each active user; an aggregation in the recommendation phase [21], [22], in the
subsequently, items not consumed by the active user that has prediction phase [23], [24] or in the neighbors obtainment
been highly valued or consumed by its neighborhood are rec- phase [39]. It is also possible to establish a virtual user that
ommended. The previous approach is classified as memory- represents the group [25], [26], b) If we opt for a model-
based [1], [9]: information to recommend is obtained directly based approach, we find that most of the models used in
from the data. The explained process is called user-based; it recommendation to individual users are not generalizable for
is also possible to carry out an item-based recommendation, recommendation to groups of users. Researchers adapt the
obtaining neighborhood sets of each item. existing models [27], [28], or they create new approaches
Currently, collaborative filtering RS are usually designed [29], [30], c) The data sparsity negative effect is greater when
by using the model-based [1], [9] approach: A model is there are more restrictions (recommendation to groups of
created from the data, and subsequently recommendations users) than when there are less restrictions (recommendation
are obtained from the model. The RS most used model- to individual users), and d) Methods with which results are
based method is the Matrix Factorization (MF) [6], [12]: measured are less universal in recommendations to groups of
The sparse ratings matrix is compressed into two dense factor users: recommendation policies, sizes and distributions of the
matrices (one matrix containing the users information and groups, etc.
another matrix containing the items information). One of the Response times in the recommendation to groups of
matrices has users x factors size, and the other matrix has users is more critical than in the individual recommenda-
items x factors size. The number of factors is usually small tions because it is necessary to perform additional opera-
(10 to 40), and then the size of each of the two matrices is tions. When the memory-based approach is used, the sooner
much smaller than the size of the original information (users the aggregation is made, the faster the results will be
x items). This compressed information contains the essence obtained [25]. Model-based approaches usually generate
of the original information, coded in factors that are called results very quickly; on the other hand, a great learning time
hidden because its meaning (the concept they encode) is not is required to generate the model.
known. The prediction and recommendation process from Fig. 1 shows the different options available to implement
this model usually improves the quality obtained through the recommendation to groups of users (when using the
memory-based approaches. memory-based approach). Gray boxes show the usual phases
Finally, it is important to highlight two important concepts: of a recommendation made with the KNN algorithm. The
1) The more information a RS gathers, the better results it four options (a to d), which flow horizontally, represent the
can provide; that’s why hybrid RS [13] are usually designed strategies followed to make recommendations to groups of
in commercial RS (typically: collaborative + content + users. Option a) Merged recommendations, uses the lists of
demographic), and 2) There are model-based RS different items recommended to each user of the group and generates
FIGURE 1. KNN-based approaches for recommending to groups of users. Figure taken from [25] has been extended.
a unified list. Option b) Merged predictions uses the most taken from the concept used in [31], where the singularity
convenient predictions lists for each user in the group and of the ratings is exploited: the ‘‘singular’’ ratings are granted
generates a unified list of predictions (there are more predic- more important than the ‘‘usual’’ ratings; e.g. in a movie
tions than recommendations). Option c) Related neighbors, that receives a huge amount of negative ratings (or absence
starts from the neighborhood of each user of the group and of ratings), a positive rating from a user is very relevant to
generates a neighborhood that represents the entire group. find her neighborhood. The singularity concept has generated
Finally, option d) Related users, creates a virtual user that important accuracy improvements when it has been applied to
represents all the group’s users: from the virtual user a tradi- several RS open datasets.
tional individual recommendation can be made. An important principle supporting the proposed similarity
measure to recommend to groups of users is the generaliza-
C. PROPOSED APPROACH FOR RECOMMENDATION tion of the singularity concept. The underlying idea is the
TO GROUPS following: we know that two users are similar if they present
In section III, the proposed method for making recommen- singular ratings in common; e.g.: both users usually listen to
dations to groups of users is explained in detail. In this an unpopular and poorly rated song. Using the same reason-
subsection, the main ideas are presented, and the decisions ing: if a group of users shares unusual ratings with a neighbor
taken are motivated. The first decision was to choose between candidate, it is very likely that she is a suitable neighbor
a model-based or a memory-based method. Our experience candidate. The condition of singularity is less likely to occur
covers both fields: [25] provides a memory-based approach, in the recommendation to groups (e.g.: a group of friends who
while [27] provides a model-based approach. The model like a very unpopular movie). This lower probability helps us
developed in [27] does not admit significant improvements, to choose neighbors that fit very well with the preferences of
whereas the solution provided in [25] does have a promising the group. In summary: the concept of singularity fits well
route to improve results, both from [25] and [27]. with the particularities of the recommendation to groups of
Published approaches usually implement aggregation of users.
neighborhoods, or predictions and recommendations merging The rest of the paper is structured as follows: section II
(options b, c and d in Fig. 1). The solution proposed in [25] summarizes the related work in the recommendation to group
provides an innovative metric (UGSM) that allows you to find of users’ sub-field. Section III defines and explains the pro-
the existing similarity between any user of the RS and the posed method. Section IV exposes the design of experi-
group of users to which you want to recommend: option a) ments, section V shows the results obtained and discuss them.
in Fig. 1. The proposed method is simple and also achieves Section VI provides the conclusions of the paper and suggests
significantly better execution times than methods based on future works. Finally, we list references, many of them of very
options b), c), d) in Fig. 1. We test its effectiveness by recent publications.
comparing it with the UGSM itself, with the model-based
method from [27] and with several baselines of very recent II. RELATED WORK
publication. The CF-based recommendation to groups field presents a
The fundamental idea on which the improvement of the wide variety in the way it has been focused. Mainly it
proposed method is based, with respect to UGSM, has been can be classified as follows: a) KNN-based approaches,
recommendation to groups approaches are proposed in [27]; 5) OTHER MODELS AND METHODS
they explain three original methods to map the group of users Using ontological concepts, Ben Ahmed et al. [52] pro-
to the latent factor space, and they compare the proposed vide a semantic multidimensional group recommendation.
methods in three different scenarios: when the group size is Meta-learning techniques and aggregation strategies [53] are
small, medium and large. Social information is used in [43]; used in order to select group recommendation strategies.
this work combines a matrix factorization model to estimate Using stochastic methods, [54] paper makes recommen-
unknown ratings with a social network analysis to evidence dations based on link-structure analysis in a probabilis-
possible social influence. tic manner, and they provide group recommendations in
social media systems. Knowledge flow mining [55] has
2) GRAPH APPROACHES been used as model to make group recommendations for
An original group recommendation approach addresses the task-based groups. A Hybrid model, merging content-based,
problem of generating and then recommending an artworks collaborative filtering, and knowledge-based solution for
sequence for a group of visitors within a museum [29]. They groups of users has been used in [56]. Social Networks
define the museum artworks as a graph. The recommendation based models play an important role in recommendation
system suggests a route that maximizes visitors’ satisfaction, to groups of users when social information is included in
recommending to individual users and then aggregate the datasets, [57] propose a social-aware group recommendation
recommendations in a single group recommendation (merged framework that jointly utilizes both social relationships and
recommendations). A random walk with restart method is social behaviors to infer a group’s preference, and also to
proposed in [44]; they represent the relationships among model the tolerance and altruism characteristics of group
users, groups, and items as a tripartite graph. Based on members.
the tripartite graph it is possible to predict the relevance
degrees between groups and unrated items by comprehen-
sively detecting their relationships. The method from [45] III. PROPOSED METHOD
models information with a heterogeneous graph and con- This section explains the proposed method for recommenda-
siders the recommendation problem as a query-dependent tion to groups of users. In the first subsection, we explain
node proximity problem. They propose a general graph-based previous concepts which are directly related to the proposed
model, called HeteRS, to recommend on Event-based social design. In the second subsection, we expose the general
networks. principles on which our recommendation to groups similarity
measure is based. The third subsection details the mathe-
3) FUZZY TECHNOLOGIES matical formalism and relates it to the explained design.
A new method for group decision making using group recom- Finally, an illustrative running example based on a datatoy
mendations based on interval fuzzy preference relations and is provided.
consistency matrices is presented in [30]. Cheng et al. [46]
propose a new autocratic decision-making method using
group recommendations based on ranking interval type-2 A. PREVIOUS CONCEPTS
fuzzy sets. A group decision making with multi-granular The proposed similarity measure uses the singularity concept
hesitant fuzzy linguistic information is presented in [47]; [31] and the combination of non-numerical and numerical
authors propose a model-based decision-making method. information concept [58]. The singularity concept hypothesis
In [48] it is considered using fuzzy logic for mod- is that CF results can be improved by taking contextual infor-
eling student clusters. As the representation of each mation, drawn from the entire body of users. The singularity
group, they assume fuzzy numbers connected with learner main idea lies in the fact that the contribution of an item to
attributes. the similarity assigned to two users ought not to be considered
as absolute (which is what happens with traditional metrics),
4) CLUSTERING PREPROCESSING METHODS but rather as relative to the vote awarded to this item by
Group recommender systems can be designed to the rest of the users in the system. Following the singularity
automatically detect groups of users by clustering them [49]; reasoning expressed through to its extremes, if all of the
the number of cluster respect a constraint on the maxi- users who have voted for an item have cast the same vote,
mum number of recommendation lists that can be produced. it is difficult to consider this item as a factor of similarity
An extensive model for group recommendations [50] exploits between two users. In the opposite case scenario, if only
recommendations for items that similar users to the group two users have voted differently to the rest for one item, this
members liked in the past. They make a pre-partition of represents a very great singularity which should be translated
users into clusters and they use the cluster members for into a very great similarity for this item. Equations from [31]
recommendations. The clustering process [51] usually make define SPi as the singularity of the relevant votes concerning
recommendations for users produced with respect to the the item i, and SNi as the singularity of the non-relevant
preferences of their cluster members without extensively votes concerning the item i. Relevant votes in MovieLens
searching for similar users in the whole user base. are 4,5, whereas non-relevant votes are 1,2,3. SPi and SNi are
FIGURE 2. Information sources used to design the proposed method. FIGURE 3. Proportion of matching items and item unpopularity.
defined as:
in the accuracy obtained when Jaccard is applied to
#{u ∈ U |ru,i ∈ {4, 5}} the KNN-based CF RS. Our proposed SMGU similarity
SPi = 1 −
#U measure extends the JMSD Jaccard behavior in two
i #{u ∈ U |ru,i ∈ {1, 2, 3}} aspects:
SN = 1 −
#U
a) Jaccard extension to groups of users: Jaccard,
Where U is the set of users and ru,i is the vote cast for user applied to two users, can be defined as the ratio
u to the item i. between the number of common cast ratings and
The combination of non-numerical and numerical informa- the total number of cast ratings. The number of
tion concept was proposed in [58]. Authors present a metric common cast ratings can be obtained using the set
which combines the numerical information of the votes with intersection operation, whereas the total number
independent information from those values, based on the of cast ratings can be obtained using the set union
proportions of the common and uncommon votes between operation. When a group of users is involved,
each pair of users. The numerical information is processed union and intersection operations can be made as
using the Mean Squared Differences (MSD) similarity mea- shown in the top of Fig. 3: each neighbor candi-
sure, whereas Jaccard is used to deal with the non-numerical date rating is compared with the corresponding
information. The resulting metric JMSD is simple, it provides group rating.
good accuracy and performance, and it has been highly cited: b) Incorporation of the item unpopularity: SGMU
JMSD(u,v) = Jaccard(u,v) * (1-MSD(u,v)). incorporates the concept ‘‘singularity of items’’
[31], expressed as the absence of popularity of an
B. METHOD DESIGN item for the users. The idea behind this concept is
Starting from the concepts shown in the previous section, that the most singular items provide better infor-
we will show the constitutive elements of the proposed simi- mation to determine the similarity between two
larity measure for groups of users. We are going to make use users. By way of example, if two users agree on
of Fig. 2; our metric is constituted from three fundamental a film that has not been widely voted and known,
elements: a) Numerical information (ratings), contemplated it is very likely that they will have similar tastes
in element 5 of Fig. 2, b) Non-numerical (structural) informa- among themselves, and different to the tastes of
tion, contemplated in element 1 of Fig. 2, and c) Weightings other users. If two users agree on a well-known
of the importance of the numerical information, contemplated movie, such as ‘‘Avatar’’, this information will
in elements 2, 3 and 4 of Fig. 2. We will call SMGU to the not be enough to match both users, since there
proposed similarity measure (Singularity-based Measure for will be a large number of users who are also
Groups of Users). paired based on their assessment of that popular
1) Proportion of matching items: Jaccard measure pro- item. Following this reasoning, item 1 from the
vides the existing proportion of votes cast by two users: bottom of Fig. 3 is more unpopular (less voted)
the active user and each neighbor candidate. JMSD sim- than item 2 because item 1 contains less ratings
ilarity measure [58] shows a significant improvement than item 2. Our proposed similarity measure will
respect to the rating of the neighbor user u, for for item i2 : it is the only rating ‘‘zero’’ of the nine votes
item i. cast in this item. The opposite situation can be observed in
P 2 item i11 : any of its three ‘‘one’’ votes gets a null singularity,
g∈Gi (rg,i − ru,i )
1G,u,i = , ∀ ru,i 6 = • (6) due to the absence of diversity in the votes. Section 3 in
#Gi Table 1 shows the group G singularity sG,i . This singularity is
While step 1 provides the non-numerical (structural) infor- computed using the geometric mean of the users’ singularities
mation, combining steps 2 to 5 into a single equation pro- in the group; in our running example, the highest sG,i value
vides the numerical information. Let’s define yG,u as the corresponds to the highest singularity value su,i from the
similarity between group G and user u, based on the numer- users of the group: s2,1 = 0.60. Section 4 in Table 1 just
ical information of the ratings. What we do is to use Mean shows the number of users in G who have voted item i; e.g.:
Squared Differences, weighting its values with three param- i12 = 2, since u1 and u3 have voted it. We can see that,
eters: a) candidate to neighbor singularity for each item, b) in this example, item i7 is particularly relevant to the SMGU
group singularity for each item, and c) number of votes in the proposed similarity measure, since it has high values of su,i ,
group for each item: sG,i and #Gi .
Section 5 shows results of the mean squared differences
i∈{IG ∩Iu } su,i sG,i #Gi 1G,u,i
P
yG,u = 1 − P (7) computed according to the equation 1G,u,i . As expected,
i∈{IG ∩Iu } su,i sG,i #Gi users u6 and u8 provide the smallest error values, whereas
Finally, the proposed SMGU similarity measure combines users u5 and u7 report the biggest errors. Combining error
non-numerical information from (2) and numerical informa- values with singularity parameters we obtain the numeri-
tion from (7). α ∈ [0, 1] is a parameter that swings the impor- cal factor yG,u . Finally, combining yG,u with the structural
tance of non-numerical information of ratings vs. numerical information xG,u , and applying the α factor we obtain the
information. SMGU similarity measure final results; in all cases: structural,
numerical and final results, the expected neighbors u6 and u8
α
SMGU (G, u) = xG,u y1−α
G,u (8) have been chosen, whereas u5 and u7 have been dismissed.
#TPλG
cross-validation. Table 3 contains the main parameters of recallGλ = (10)
each dataset, chosen for the experiments execution. #TGλ
B. QUALITY MEASURES Where TP, FP and T denote the true positive, false positive
In order to evaluate the quality of recommendations to groups and expected recommendations sets, respectively:
of users, we define precision and recall for the group G as:
FPλG = i ∈ LG |aG,i < λ
(11)
#TPλG TPλG = i ∈ LG |aG,i ≥ λ
(12)
precisionλG = (9)
TPλG ∪ FPλG TGλ = i ∈ I |aG,i ≥ λ
# (13)
TABLE 2. Main properties of the datasets used in the experiments. TABLE 4. Tested model parameters.
L̂uG = i ∈ LG |r̂u,i 6 = •
(17) D. MODEL PERFORMANCE
To measure the performance of the proposed model we
We define R−MAEuG as the prediction error for the user u will compare it with state of the art group recommenda-
w.r.t. the item recommended to the group G: tion methods. The baselines selected for this comparison
are UGSM [25], MFGU [27], C&P [32], RWR − M [20]
P
i∈L̂uG r̂u,i − pG,i
G and RANK [22]. Some of these recommendations methods
R−MAEu = (18)
#L̂uG require different parameters to work. We have configured that
parameters in order to maximize the quality of the recommen-
We define R−MAE as the averaged error of the predictions
dation in each dataset. Table 6 contains the parameters of each
performed to the group G:
model in each experiment.
R−MAEuG We will evaluate these methods using the previously
P
R−MAEG = u∈G (19) defined quality measures: precision, recall and R−MAE.
#G
39754 VOLUME 6, 2018
F. Ortega et al.: Recommendation to Groups of Users Using the Singularities Concept
FIGURE 5. α parameter and number of neighbors impact on the quality results obtained by using the proposed method.
The experiment has been carried out for two group sizes: recommendation results obtained using the proposed method,
a) from 2 to 5 users, to test the performance of each recom- then compares them with those offered by processing the
mendation method on small groups of users (families, cou- baselines, and 3) Discusses the observed main strengths and
ples, small group of friends, etc.), and b) from 6 to 10 users, drawbacks.
to test the performance of each recommendation method over
large group of users (teams, bands, large group of friends, A. PARAMETERS SETTING
etc.). Furthermore, we have tested the experiment for mul- Before testing the proposed method it is necessary to deter-
tiple λ values (1.0, 0.75 and 0.5) in order to measure the mine the alpha value that we will use in equation (8). That
overall satisfaction of the group within the recommendations is, we must choose an α value that weights the importance of
performed. Each measure has been evaluated for different the structural (non-numerical) information of the votes with
numbers of recommendations (from 5 to 20). respect to the numerical information. The lower the value of
α, the greater weight we give to the numerical information
V. RESULTS and the less weight we give to the non-numerical information.
This section is divided into three sub-sections: 1) Explains Fig. 5 implements Table 4 and summarizes the combined
the experiments carried out to adjust the parameters involved impact that the α values and the numbers of neighbors pro-
in the proposed method, 2) Shows the prediction and duce on each selected quality measure and for each of the
considered datasets. In all the graphs the best results are similarity measure SMGU achieves the best prediction results
coded using the yellow color (low values in R_MAE and in all cases. SMGU and MFGU behave very well when
high values in Precision and Recall). The main conclusions applied to large datasets. 3) In general, the quality of the
from the results shown in Fig. 5 are: 1) The optimal α and prediction increases as the size of the group decreases: this
number of neighbors values depend, to a large extent, on the is the expected result, due to the greater diversity of the users
nature of the data; that is: the dataset we are using, 2) The that make up the larger groups.
structural (non-numerical) information helps to improve the Once the superiority of the proposed method, in predic-
quality results: when α is zero the quality results get worse, tion quality, has been verified, we carry out the experiments
and 3) Generally, it is appropriate to assign low α values: we that measure the recommendation quality; for this, we test
should give more importance to the numerical information precision@λ and recall@λ, and we establish the following λ
than to the non-numerical information. values: 0.5, 0.75, 1.0. As an example, the value 0.75 means
The values of α and number of neighbors exposed that there is a proportion of 75% of users, belonging to each
in Table 5 have been selected based on the results shown group, that like the recommended items. The recommenda-
in Fig. 5. It is interesting to observe, as expected, how the tion quality measures have been tested using all the consid-
size of the dataset influences the appropriate value of the ered baselines for groups of users: UGSM [25], MFGU [27],
number of neighbors parameter; e.g.: in the largest dataset C&P [32], RWR − M [20] and RANK [22]. Experiments
(MovieLens) the impact is very small, whereas in the smallest parameters are shown in Table 6. Fig. 7 shows the recom-
dataset (FilmTrust) there is a big difference in the quality of mendation quality results. Each function in Fig. 7 is obtained
the results obtained by applying different neighborhood sizes. through the results of ten recommendations. Experiment has
It is also very interesting to observe, in FilmTrust, the fol- been performed for small groups from 2 to 5 users (left side
lowing behavior: there is a tendency to obtain the best quality of the figure) and large groups from 6 to 10 users (right side
results when establishing a direct correlation between α and of the figure).
number of neighbors. That is, as we increase the size of Fig. 7 shows the proposed similarity measure superiority
the neighborhood, we must also increase the α value. This for the majority of the Prediction quality results (the higher
means that the best neighbors (the first ones, the most similar the value, the better the result). Specifically, in the larger
ones) are best selected using numerical information: low α. dataset: MovieLens, the improvement of results offered by
However, the more ‘‘forced’’ neighbors (the last ones, the less SMGU is clearly visible. MFGU is the baseline that, in large
similar ones) are better selected using structural information: datasets, best competes with the proposed group-based simi-
high α. Our proposed method can compensate this effect larity measure; this is because, being a model-based method,
by raising or by decreasing the value of α. The R_MAE its design is better suited to large volumes of data. The pro-
and Precision graphs, in MovieLens, also show this trend, posed similarity measure does not always get the best results
although in a less marked way due to the much larger size from Recall; specifically, in the MovieLens dataset, the Rank
of this dataset: It would be necessary to use a greater number baseline method improves to SMGU . As expected, the small
of neighbors to appreciate this trend with greater intensity. groups results tend to be better than the large groups results,
due to the greater variation of users in large groups.
B. COMPARATIVE RESULTS The set of experiments performed for each dataset, for each
First of all, we will show the prediction quality results baseline and for the proposed similarity measure have been
making use of the quality measure R_MAE ‘‘Mean abso- replicated in each of the quality measures: precision@0.5,
lute prediction error of the recommended items’’. Fig. 6 recall@0.5, precision@1.0 and recall@1.0. The qualita-
shows the obtained results for the three tested datasets: tive results are similar to those explained in this section
MovieLens, FilmTrust and BookCrossing. The used baselines (precision@0.75 and recall@0.75), so the resulting fig-
are: UGSM [25], MFGU [27] and C&P [32]. Baselines ures are not included. These figures are stored online,
RWR−M [20] and RANK [22] are not designed to predict; we in the journal site, as ‘‘Additional Material’’. The quantitative
will use them to test the recommendation quality measures results show the expected behavior: the more demanding the
Prediction and Recall. Experiments parameters are shown requested quality is (i.e.: the higher the λ value), the lower the
in Table 6. Experiments have been carried out for different precision and recall results obtained.
group sizes: left side correspond with small groups of users The proposed SMGU mostly improves the results of
(groups containing from two to five users) and right side the best baseline MFGU : R_MAE, Precision and Recall.
correspond with large groups of users (groups containing However, the Recall of SMGU is not always the most appro-
from six to ten users). priate (Fig. 7). In short, our proposed method obtains bet-
The most relevant conclusions obtained from Fig. 6 are: ter Precision than Recall. The question that arises is the
1) The quality offered by the baseline similarity measures importance and meaning of each of these recommendation
vary depending on the dataset used and the number of users quality measures. Simplifying, we can associate Precision
of the groups: in some cases, UGSM exceeds MFGU , while with an absolute quality value, and Recall with a relative
on other occasions the opposite occurs. C&P tends to be quality value: the first one measures the success with respect
the baseline that provides the worst results. 2) The proposed to a fixed number of recommendations, while the second
FIGURE 6. Mean absolute prediction error of the recommended items in MovieLens, FilmTrust and BookCrossing datasets. Tested for small groups (left
side) and large groups (right side). Proposed method: SMGU . Baselines: UGSM [25], MFGU [27] and C &P [32]. Experiments parameters are shown
in Table 6. The lower the value, the better the result.
one measures the success with respect to the number of best Recall results when the number of recommendations
cases in which it is appropriate to recommend. Precision is low. When the number of recommendations is not low,
focuses on quantities (accuracy, effectiveness) while Recall Rank provides better Recall results. MGFU does not improve
focuses on qualities (it rewards complicated hits). In short, our proposed similarity measure neither in Precision nor in
we can determine that the proposed SMGU method and Recall.
the main MFGU baseline are accurate (a large number
of recommendations are correct), although they are not C. DISCUSSION
always the ones that provide the most complex recommen- The SMGU proposed similarity measure to recommend to
dations; e.g.: by using MovieLens, Rank gets good Recall groups of users has important advantages, but also some
results. drawback that should be analyzed and given the importance
Fig. 7 functions show Precision and Recall results when it deserves. The main advantage of SMGU is its superiority
the number of recommendations range from two (left side of in prediction and also in recommendation, measured, respec-
each function) to ten (right side of each function). We can tively, with the R_MAE and Precision quality measures.
observe that, for the MovieLens dataset, SMGU provides the Additional advantages of SMGU are: a) It maintains the best
FIGURE 7. Recommendation results: Precision and Recall with a proportion of 75% of users belonging to the group that liked the recommended items.
MovieLens, FilmTrust and BookCrossing datasets. Number of recommendations: 10. Tested for small groups (left side) and large groups (right side).
Proposed method: SMGU . Baselines: UGSM [25], MFGU [27], C &P [32], RWR − M [20] and RANK [22]. Experiments parameters are shown in Table 6. The
higher the value, the better the result (both precision & recall).
results for different sizes of the groups, b) It improves the that the concept of singularity fits well with the particularities
baselines results when applied to datasets with different sizes of the recommendation to groups of users.
and diverse sparsities, and c) It recommends adequately both SMGU is a memory-based similarity measure. Most of
when a small number of recommendations is required and the published papers offer memory-based solutions to the
when a larger number is needed. recommendations to groups of users, but the MFGU method
An interesting feature of SMGU is its ability to weight the is model-based. The memory-based versus model-based dis-
importance of the numerical information of ratings with the cussion in the recommendation to groups of users is the same
non-numerical information. This weighting mechanism also as in the recommendation to individual users: each approach
serves to adapt to the number of neighbors selected in the has its advantages and disadvantages. However, while in rec-
CF recommendation process and to maintain high levels of ommendation to individual users the model-based methods
accuracy. The concept of singularity is used intensively in usually offer better accuracy, in our case the memory-based
SMGU . This concept has proven to provide good results in the SMGU similarity measure improves the results of the MFGU
recommendation to individual users. In this paper, it is shown method.
VI. CONCLUSIONS [8] Z. Yang, B. Wu, K. Zheng, X. Wang, and L. Lei, ‘‘A survey of collaborative
Recommendation to groups of users is a challenging research filtering-based recommender systems for mobile Internet applications,’’
IEEE Access, vol. 4, pp. 3273–3287, 2016.
open area. It is important to find suitable information sources, [9] Y. Shi, M. Larson, and A. Hanjalic, ‘‘Collaborative filtering beyond the
coming from ratings, in order to improve accuracy. This paper user-item matrix: A survey of the state of the art and future challenges,’’
focuses on the assumption that the singularity of the ratings ACM Comput. Surv., vol. 47, pp. 3:1–3:45, May 2014.
[10] J. Bobadilla, F. Ortega, A. Hernando, and G. G. de Rivera, ‘‘A similarity
cast by the users of the group is a relevant information for metric designed to speed up, using hardware, the recommender systems
finding suitable neighbors. Results confirm this hipothesis: k-nearest neighbors algorithm,’’ Knowl.-Based Syst., vol. 51, pp. 27–34,
accuracy has been improved both for prediction results and Oct. 2013.
[11] J. Bobadilla, F. Ortega, A. Hernando, and Á. Arroyo, ‘‘A balanced memory-
for recommendation results. based collaborative filtering similarity measure,’’ Int. J. Intell. Syst.,
The proposed similarity measure to recommend to groups vol. 27, no. 10, pp. 939–946, 2012.
of users combine structural information with singularities- [12] X. Guan, C. T. Li, and Y. Guan, ‘‘Matrix factorization with rating com-
pletion: An enhanced SVD model for collaborative filtering recommender
weighted numerical information. Singularities weights systems,’’ IEEE Access, vol. 5, pp. 27668–27678, 2017.
improve accuracy just using the regular numeric ratings [13] T. K. Paradarami, N. D. Bastian, and J. L. Wightman, ‘‘A hybrid rec-
values. For each processed item, we provide three different ommender system using artificial neural networks,’’ Expert Syst. Appl.,
vol. 83, pp. 300–313, 2017.
parameters to weight numerical information: 1) Singularity [14] S. P. Perumal, K. Arputharaj, and G. Sannasi, ‘‘Fuzzy family tree similarity
of the group ratings, 2) Singularity of each neighbor rating, based effective e-learning recommender system,’’ in Proc. 8th Int. Conf.
and 3) Number of users in the group rating the item. Analo- Adv. Comput. (ICoAC), Jan. 2017, pp. 146–150.
[15] T. Horváth and A. C. P. L. F. de Carvalho, ‘‘Evolutionary computing in
gously, non-numerical information is weighted attending to recommender systems: A review of recent research,’’ Natural Comput.,
the popularity of items. vol. 16, pp. 441–462, Sep. 2017.
Results show a large accuracy improvement of the pro- [16] A. Hernando, J. Bobadilla, and F. Ortega, ‘‘A non negative matrix factoriza-
tion for collaborative filtering recommender systems based on a Bayesian
posed method compared to the KNN baselines, both for probabilistic model,’’ Knowl.-Based Syst., vol. 97, pp. 188–202, Apr. 2016.
prediction and for recommendation. Compared to the matrix [17] J. Bobadilla, R. Bojorque, A. H. Esteban, and R. Hurtado, ‘‘Recommender
factorization-based baseline, the proposed method shows systems clustering using Bayesian non negative matrix factorization,’’
IEEE Access, vol. 6, pp. 3549–3564, 2018.
more adjusted improvements. Therefore, our method is pre- [18] J. Masthoff, Group Modeling: Selecting a Sequence of Television Items
sented as the currently most appropriate to meet diverse to Suit a Group of Viewers. Dordrecht, The Netherlands: Springer, 2004,
objectives and restrictions: a) KNN-based recommender sys- pp. 93–141.
[19] J. Masthoff, Group Recommender Systems: Combining Individual Models.
tems for groups of users, b) Systems providing continuously Boston, MA, USA: Springer, 2011, pp. 677–702.
updated recommendations, c) Explanation of recommenda- [20] S. Feng and J. Cao, ‘‘Improving group recommendations via detecting
tions to the group due to the superiority of KNN approaches comprehensive correlative information,’’ Multimedia Tools Appl., vol. 76,
pp. 1355–1377, Jan. 2017.
to explain recommendations against model based ones. [21] M. O’Connor, D. Cosley, J. A. Konstan, and J. Riedl, PolyLens: A Recom-
As future works we propose: a) To improve the proposed mender System for Groups of Users, Dordrecht, The Netherlands: Springer,
similarity measure by incorporating new elements of simi- 2001, pp. 199–218.
[22] L. Baltrunas, T. Makcinskas, and F. Ricci, ‘‘Group recommendations with
larity processing between groups and neighbors candidates. rank aggregation and collaborative filtering,’’ in Proc. 4th ACM Conf.
b) Explaining recommendations to the group according to Recommender Syst. (RecSys), New York, NY, USA, 2010, pp. 119–126.
both the similarity values and the singularity weights, c) To [23] I. A. Christensen and S. Schiaffino, ‘‘Entertainment recommender systems
for group of users,’’ Expert Syst. Appl., vol. 38, no. 11, pp. 14127–14135,
study beyond accuracy quality results (such as: novelty, diver- 2011.
sity or reliability) for each member of the group, and d) To [24] S. Berkovsky and J. Freyne, ‘‘Group-based recipe recommendations: Anal-
incorporate singularity weights into the matrix factorization ysis of data aggregation strategies,’’ in Proc. 4th ACM Conf. Recommender
Syst. (RecSys), New York, NY, USA, 2010, pp. 111–118.
process to improve matrix factorization-based recommenda-
[25] F. Ortega, J. Bobadilla, A. Hernando, and A. Gutiérrez, ‘‘Incorporating
tion to groups of users. group recommendations to recommender systems: Alternatives and per-
formance,’’ Inf. Process. Manage., vol. 49, no. 4, pp. 895–901, 2013.
[26] K. McCarthy, M. Salamó, L. Coyle, L. McGinty, B. Smyth, and P. Nixon,
REFERENCES
‘‘CATS: A synchronous approach to collaborative group recommenda-
[1] J. Bobadilla, F. Ortega, A. Hernando, and A. Gutiérrez, ‘‘Recommender tion,’’ in Proc. Florida Artif. Intell. Res. Soc. Conf. (FLAIRS), 2006,
systems survey,’’ Knowl.-Based Syst., vol. 46, pp. 109–132, Jul. 2013. pp. 86–91.
[2] J. Lu, D. Wu, M. Mao, W. Wang, and G. Zhang, ‘‘Recommender sys- [27] F. Ortega, A. Hernando, J. Bobadilla, and J. H. Kang, ‘‘Recommending
tem application developments: A survey,’’ Decis. Support Syst., vol. 74, items to group of users using matrix factorization based collaborative
pp. 12–32, Jun. 2015. filtering,’’ Inf. Sci., vol. 345, pp. 313–324, Jun. 2016.
[3] M.-L. Wu, C.-H. Chang, and R.-Z. Liu, ‘‘Integrating content-based fil- [28] X. Hu, X. Meng, and L. Wang, ‘‘Svd-based group recommendation
tering with collaborative filtering using co-clustering with augmented approaches: An experimental study of moviepilot,’’ in Proc. 2nd Challenge
matrices,’’ Expert Syst. Appl., vol. 41, no. 6, pp. 2754–2761, 2014. Context-Aware Movie Recommendation (CAMRa), New York, NY, USA,
[4] M. Y. H. Al-Shamri, ‘‘User profiling approaches for demographic recom- 2011, pp. 23–28.
mender systems,’’ Knowl.-Based Syst., vol. 100, pp. 175–187, May 2016. [29] S. Rossi, F. Barile, C. Galdi, and L. Russo, ‘‘Recommendation in museums:
[5] J. Tang, X. Hu, and H. Liu, ‘‘Social recommendation: A review,’’ Social Paths, sequences, and group satisfaction maximization,’’ Multimedia Tools
Netw. Anal. Mining, vol. 3, no. 4, pp. 1113–1133, 2013. Appl., vol. 76, pp. 26031–26055, Dec. 2017.
[6] J. Yu, M. Gao, W. Rong, Y. Song, and Q. Xiong, ‘‘A social recommender [30] S.-M. Chen, S.-H. Cheng, and T.-E. Lin, ‘‘Group decision making systems
based on factorization and distance metric learning,’’ IEEE Access, vol. 5, using group recommendations based on interval fuzzy preference relations
pp. 21557–21566, 2017. and consistency matrices,’’ Inf. Sci., vol. 298, pp. 555–567, Mar. 2015.
[7] N. M. Villegas, C. Sánchez, J. Díaz-Cely, and G. Tamura, ‘‘Characteriz- [31] J. Bobadilla, F. Ortega, and A. Hernando, ‘‘A collaborative filtering simi-
ing context-aware recommender systems: A systematic literature review,’’ larity measure based on singularities,’’ Inf. Process. Manage., vol. 48, no. 2,
Knowl.-Based Syst., vol. 140, pp. 173–200, Jan. 2018. pp. 204–217, 2012.
[32] L. Boratto, S. Carta, and G. Fenu, ‘‘Investigating the role of the rating [54] H.-N. Kim and A. E. Saddik, ‘‘A stochastic approach to group recommen-
prediction task in granularity-based group recommender systems and big dations in social media systems,’’ Inf. Syst., vol. 50, pp. 76–93, Jun. 2015.
data scenarios,’’ Inf. Sci., vol. 378, pp. 424–443, Feb. 2017. [55] C.-H. Lai, ‘‘Applying knowledge flow mining to group recommendation
[33] S. Berkovsky, J. Freyne, and M. Coombe, ‘‘Aggregation trade offs methods for task-based groups,’’ J. Assoc. Inf. Sci. Technol., vol. 66, no. 3,
in family based recommendations,’’ in Advances in Artificial Intelli- pp. 545–563, 2015.
gence, A. Nicholson and X. Li, Eds. Berlin, Germany: Springer, 2009, [56] T. D. Pessemier, J. Dhondt, and L. Martens, ‘‘Hybrid group recommenda-
pp. 646–655. tions for a travel service,’’ Multimedia Tools Appl., vol. 76, pp. 2787–2811,
[34] J. Freyne and B. Smyth, ‘‘Cooperating search communities,’’ in Adaptive Jan. 2017.
Hypermedia and Adaptive Web-Based Systems, V. P. Wade, H. Ashman, [57] L. Sun, X. Wang, Z. Wang, H. Zhao, and W. Zhu, ‘‘Social-aware video rec-
and B. Smyth, Eds. Berlin, Germany: Springer, 2006, pp. 101–110. ommendation for online social groups,’’ IEEE Trans. Multimedia, vol. 19,
[35] I. Garcia, L. Sebastia, and E. Onaindia, ‘‘On the design of individual and no. 3, pp. 609–618, Mar. 2017.
group recommender systems for tourism,’’ Expert Syst. Appl., vol. 38, [58] J. Bobadilla, F. Serradilla, and J. Bernal, ‘‘A new collaborative filtering
no. 6, pp. 7683–7692, 2011. metric that improves the behavior of recommender systems,’’ Knowl.-
[36] M. Stettinger and A. Felfernig, ‘‘Choicla: Intelligent decision support for Based Syst., vol. 23, no. 6, pp. 520–528, 2010.
groups of users in the context of personnel decisions,’’ in Proc. ACM [59] F. M. Harper and J. A. Konstan, ‘‘The movielens datasets: History and con-
RecSys IntRS Workshop, 2014, pp. 28–32. text,’’ ACM Trans. Interact. Intell. Syst., vol. 5, pp. 19-1–19-19, Dec. 2015.
[37] Y. Chen and P. Pu, ‘‘Cofeel: Using emotions to enhance social interaction in [60] Z. Zaier, R. Godin, and L. Faucher, ‘‘Evaluating recommender systems,’’ in
group recommender systems,’’ in Proc. Workshop Tools Technol. Emotion- Proc. Int. Conf. Automated Solutions Cross Media Content Multi-Channel
Awareness Comput. Mediated Collaboration Learn. Alpine Rendez-Vous Distrib., Nov. 2008, pp. 211–217.
(ARV), 2013, pp. 1–2. [61] J. Golbeck and J. Hendler, ‘‘FilmTrust: Movie recommendations using
[38] J. J. Jung, ‘‘Attribute selection-based recommendation framework for trust in web-based social networks,’’ in Proc. 3rd IEEE Consum. Commun.
short-head user group: An empirical study by movielens and IMDB,’’ Netw. Conf. (CCNC), vol. 1, Jan. 2006, pp. 282–286.
Expert Syst. Appl., vol. 39, no. 4, pp. 4049–4054, 2012. [62] F. Ortega, B. Zhu, J. Bobadilla, and A. Hernando, ‘‘CF4J: Collaborative
[39] J. Bobadilla, F. Ortega, A. Hernando, and J. Bernal, ‘‘Generalization of filtering for java,’’ Knowl.-Based Syst., vol. 152, pp. 94–99, Jul. 2018.
recommender systems: Collaborative filtering extended to groups of users
and restricted to groups of items,’’ Expert Syst. Appl., vol. 39, no. 1,
pp. 172–186, 2012.
[40] Z. Yu, X. Zhou, Y. Hao, and J. Gu, ‘‘TV program recommendation for
multiple viewers based on user profile merging,’’ User Model. User-
Adapted Interact., vol. 16, no. 1, pp. 63–82, 2006.
[41] Y. Zhiwen, Z. Xingshe, and Z. Daqing, ‘‘An adaptive in-vehicle multimedia
recommender for group users,’’ in Proc. IEEE 61st Vehicular Technol.
Conf. (VTC-Spring), vol. 5, May 2005, pp. 2800–2804.
[42] Y. Liu, L. Yang, J. Sun, Y. Jiang, and J. Wang, ‘‘Collaborative matrix
factorization mechanism for group recommendation in big data-based
library systems,’’ Library Hi Tech, vol. 36, no. 3, pp. 458–481, 2018. FERNANDO ORTEGA was born in 1988. He
[43] I. Christensen and S. Schiaffino, ‘‘Matrix factorization in social group received the B.S. degree in software engineering
recommender systems,’’ in Proc. 12th Mexican Int. Conf. Artif. Intell.,
and the Ph.D. degree in computer sciences from
Nov. 2013, pp. 10–16.
the Universidad Politécnica de Madrid, Madrid,
[44] Y. Bao, H. Fang, and J. Zhang, ‘‘TopicMF: Simultaneously exploiting
in 2010 and 2015, respectively. He is currently
ratings and reviews for recommendation.,’’ in Proc. 28th AAAI Conf. Artif.
Intell., vol. 14, pp. 2–8, Jul. 2014. an Assistant Professor with the Deparment of
[45] T. A. N. Pham, X. Li, G. Cong, and Z. Zhang, ‘‘A general graph-based Engineering, U-tad: Centro Universitario de Tec-
model for recommendation in event-based social networks,’’ in Proc. IEEE nología y Arte Digital. His main research fields
31st Int. Conf. Data Eng., Apr. 2015, pp. 567–578. are machine learning, data analysis, and artificial
[46] S.-H. Cheng, S.-M. Chen, and Z.-C. Huang, ‘‘Autocratic decision making intelligence. He has published several papers in the
using group recommendations based on ranking interval type-2 fuzzy most relevant international journals. He also actively collaborates in various
sets,’’ Inf. Sci., vols. 361–362, pp. 135–161, Sep. 2016. projects with technology companies.
[47] W. Yu, Q. Zhong, and Z. Zhang, ‘‘Multi-granular hesitant fuzzy linguistic
group decision making with incomplete weight information,’’ Syst. Eng.,
vol. 38, no. 3, pp. 777–785, 2018.
[48] K. Myszkorowski and D. Zakrzewska, ‘‘Using fuzzy logic for recom-
mending groups in e-learning systems,’’ in Computational Collective Intel-
ligence. Technologies and Applications, C. Bǎdicǎ, N. T. Nguyen, and
M. Brezovan, Eds. Berlin, Germany: Springer, 2013, pp. 671–680.
[49] L. Boratto and S. Carta, ‘‘Art: Group recommendation approaches for
automatically detected groups,’’ Int. J. Mach. Learn. Cybern., vol. 6,
pp. 953–980, Dec. 2015.
[50] E. Ntoutsi, K. Stefanidis, K. Nørvåg, and H.-P. Kriegel, ‘‘Fast group
recommendations by applying user clustering,’’ in Conceptual Modeling, REMIGIO HURTADO was born in 1989. He
P. Atzeni, D. Cheung, and S. Ram, Eds. Berlin, Germany: Springer, 2012, received the B.S degree in systems engineer-
pp. 126–140. ing from the Universidad Politécnica Salesiana,
[51] I. Ntoutsi, K. Stefanidis, K. Norvag, and H.-P. Kriegel, ‘‘gRecs: A group Ecuador, in 2012, the master’s degree in informa-
recommendation system based on user clustering,’’ in Database Systems
tion and software technology from the Instituto
for Advanced Applications, S. Lee, Z. Peng, X. Zhou, Y. S. Moon,
Tecnológico y de Estudios Superiores de Monter-
R. Unland, and J. Yoo, Eds. Berlin, Germany: Springer, 2012,
pp. 299–303.
rey, Mexico, in 2014, and the master’s degree in
[52] E. Ben Ahmed, W. Tebourski, W. Ben Abdessalem Karaa, and F. Gargouri, computer science and technology from the Univer-
‘‘SMART: Semantic multidimensional group recommendations,’’ Multi- sidad Politécnica de Madrid, Spain, in 2017.
media Tools Appl., vol. 74, pp. 10419–10437, Dec. 2015. He is currently a Lecturer with the Universi-
[53] A. Zapata, V. H. Menéndez, M. E. Prieto, and C. Romero, ‘‘Evaluation and dad Politécnica Salesiana. His research interests include the recommender
selection of group recommendation strategies for collaborative searching systems and natural language processing. He is currently a member of the
of learning objects,’’ Int. J. Human-Comput. Stud., vol. 76, pp. 22–39, Research Team in Artificial Intelligence and Assistive Technology, his team
2015. is collaborating with the Universidad Politécnica de Madrid.
JESÚS BOBADILLA received the B.S. degree in RODOLFO BOJORQUE was born in 1982. He
computer science from the Universidad Politéc- received the B.S. degree in systems engineer-
nica de Madrid and the Ph.D. degree in com- ing from the Universidad Politécnica Salesiana,
puter science from the Universidad Carlos III Ecuador, in 2006, the master’s degree in secu-
de Madrid. He is currently a Lecturer with the rity information and communications technologies
Department of Applied Intelligent Systems, Uni- from Universidad Autónoma de Barcelona, Spain,
versidad Politécnica de Madrid. He is a habitual in 2015, and the master’s degree in computer sci-
author of programming languages books working ence and technology from Universidad Politécnica
with McGraw-Hill, Ra-Ma, and Alfa Omega pub- de Madrid, Spain, in 2017.
lishers. His research interests include information He is currently a Lecturer with the Universidad
retrieval, recommender systems, and speech processing. He was in charge of Politécnica Salesiana, Ecuador. His research interests includes the machine
the FilmAffinity.com Research Team, involved in the collaborative filtering learning techniques and intelligent systems. He is currently a member of
kernel of the website. He has been a Researcher with the International research team in artificial intelligence and assistive technology, his team is
Computer Science Institute, Berkeley University, Berkeley, CA, USA. collaborating with the Universidad Politécnica de Madrid.